Same to the eye. Different to the detector.
Drag to compare. Switch to the difference to see the mark. The card shows what the detector says.
Marked: the file Laidline writes, against the original.
Measured on the lossless PNGs, marked against original, 8-bit RGB. PSNR: 10 log10(255^2 / MSE) over every RGB sample. SSIM: Wang et al. 2004 on BT.601 luma (0.299 R + 0.587 G + 0.114 B), 11x11 Gaussian window with sigma 1.5, K1 0.01, K2 0.03, L 255, the mean over every full window. Pixels changed: the share of pixels where any channel differs. Largest change: the largest absolute difference in any channel, as a share of 255.
Measured on this file only. Robustness across edits is on the proof page.
The card comes from Laidline's own deployment, which accepts two issuer numbers, so its false-accept bound sits just above the one-number figure on Proof.
Three views. One photograph.
Original
The file as the model made it.
Marked
The file Laidline writes: watermark, C2PA manifest, evidence entry.
Difference ×40
The two files subtracted, amplified ×40.
Every frame is a real file. Check it yourself.
The viewer shows delivery copies, because lossy formats destroy an invisible mark. The downloads are the lossless PNGs, and the detector results come from them.
$ laidline detect marked.pngOn a fresh install the marked file reads “Signals present, unverified”, grade B. The C2PA signature reads “Valid, untrusted certificate”, because this demo signed with a development certificate. The watermark reads “Present, unverified, grade D”, and its issuer line says “Signed by Registered Issuer 1, watermark agrees; Registered to Laidline, operator of laidline.com, per directory snapshot 2026-10-21”. The grade A capture is what the deployment that made the file sees.
